Omnidirectional Dense Regression With Distortion-Free CNN Attention

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for omnidirectional dense regression, such as Fully Convolutional Networks (FCNs), fail to account for varying levels of image distortion in omnidirectional images and lack global context, while self-attention modules are not suitable due to incompatible position embedding and distortion differences across image regions.

Innovation Solution

A method and device utilizing a distortion-free CNN encoder and decoder with spherical self-attention, incorporating skip connections to handle varying distortions and compute appearance and positional attention using spherical distances, enabling effective omnidirectional image processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a conventional FCN is used for dense regression, then local feature encoding is achieved, but global context information is lost and distortion in omnidirectional images is not handled

Engineering Contradiction:

Improvedepth estimation accuracyVSAvoidglobal context information

Core Design Contradiction:

Measurement precisionVSLoss of information

Solution Approach 1:

The patent combines FCN and Transformer architectures into a hybrid model where the FCN encoder extracts local features and the Transformer decoder integrates global context information through self-attention mechanisms, resolving the contradiction between local feature encoding and global context preservation

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent introduces spherical coordinate system and spherical distance calculations to handle the curved geometry of omnidirectional images, transforming the flat 2D image processing into spherical space where distortion is naturally accounted for, thereby improving depth estimation accuracy

Inventive Principle:

Principle #14Spheroidality (Curvature)

2Loss of information

If a typical self-attention module is used, then global context information is utilized, but it is incompatible with omnidirectional image distortion and position embedding

Engineering Contradiction:

Improveglobal context informationVSAvoidcompatibility with omnidirectional space

Core Design Contradiction:

Loss of informationVSAdaptability or versatility

Solution Approach 1:

The patent applies different processing strategies to different regions of the omnidirectional image by using spherical coordinate transformations that adapt to the local distortion characteristics of each region, making the self-attention mechanism compatible with varying distortion levels across the image

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent changes the parameter space from Cartesian coordinates to spherical coordinates, transforming the position embedding and distance calculations to be compatible with omnidirectional geometry, thereby enabling the self-attention module to effectively process global context information

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12626478B2Method for omnidirectional dense regression for machine perception tasks via distortion-free CNN and spherical self-attention

Publication Date: 2026.05.12 ROBERT BOSCH GMBH

AI summary

A method and device for performing a perception task are disclosed. The method and device incorporate a dense regression model. The dense regression model advantageously incorporates a distortion-free convolution technique that is designed to accommodate and appropriately handle the varying levels of distortion in omnidirectional images across different regions. In addition to distortion-free convolution, the dense regression model further utilizes a transformer that incorporates an spherical self-attention that use distortion-free image embedding to compute an appearance attention and uses spherical distance to compute a positional attention.